Research map / Label-efficient and robust learning / Recognition and long-tail

Hierarchical Novelty Detection via Fine-Grained Evidence Allocation

ICML 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Group-robust classification

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes E-HND, an evidential loss over a flattened hierarchy with novel non-leaf classes that creates an evidence margin separating known leaf classes from novel ones for hierarchical novelty detection.

Problem

Hierarchical novelty detection (HND): detecting novel samples and assigning them to their closest known parent class in a class hierarchy; existing leave-one-out training confuses known and novel classes at test time.

Approach

Flatten the hierarchy into leaf classes plus novel non-leaf classes and train an evidential (Dirichlet) model with a KL-based loss that assigns high evidence to the known ground-truth leaf and bounded moderate evidence to the novel ground-truth parent, creating an evidence margin; base rates encode prior beliefs about hierarchy incompleteness.

Key results

• On CUB, E-HND reaches NA@50 46.18 vs 42.25 for LOO and 40.42 for DARTS; TD+E-HND reaches 46.85 NA@50 and 35.78 AUC vs TD+LOO 44.42 / 34.31.
• On Traffic, TD+E-HND achieves NA@50 52.53 and AUC 45.56, versus 50.25 / 42.86 for TD+LOO.
• Paired t-tests over 10 seeds report p-values below 1e-8 for TD+E-HND vs TD+LOO across all datasets and metrics.

Measured effects reported in the paper

Surprising findings

• Using the standard evidential losses (log, digamma, MSE) from Sensoy et al. performs worse than the proposed margin loss, and MSE loss falls to 13.15 NA@50 on CUB.
• Increasing the novel base rate cannot push novel accuracy on CUB beyond about 57%, showing a ceiling from training with known samples only.

Limitations

• Relies on frozen ResNet-101 features rather than end-to-end training, which may limit performance.
• Hyperparameters beta1 and beta2 must be chosen on a validation set and require keeping beta1 > beta2; very large beta1 degrades performance.
• Novel-class accuracy remains limited (e.g., below 60% on CUB), and the deeper hierarchy in Tiny ImageNet lowers the upper bound on N-ACC.

Experimental setup

Datasets: CUB-200-2011 TinyImageNet AWA2 Mapillary Traffic Sign Dataset ImageNet-1k TT100K

Benchmarks: CUB-200-2011 TinyImageNet AWA2 Mapillary Traffic Sign (MTSD) ImageNet-1k TT100K

Metrics: NA@50 AUC of known vs novel accuracy curve K-ACC N-ACC

Baselines: DARTS Relabel Evidential HCL LOO TD+LOO

Builds on: leave-one-out training (Lee et al., 2018) evidential deep learning (Sensoy et al., 2018) Top-Down classifiers (Lee et al., 2018)

Closest papers by meaning

• Learning State-Based Node Representations from a Class Hierarchy for Fine-Grained Open-Set Detection (ICML 2025 · similarity 0.84)
• AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing (NeurIPS 2025 · similarity 0.83)
• MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 2025 · similarity 0.83)
• Universal Novelty Detection Through Adaptive Contrastive Learning (CVPR 2024 · similarity 0.82)
• What Do You See in Common? Learning Hierarchical Prototypes over Tree-of-Life to Discover Evolutionary Traits (ICLR 2025 · similarity 0.82)
• Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.82)